Agentic AutoMLOps for Hospitals: A Self-Managing Platform for Medical AI Model Monitoring and Deployment

Abstract

Hospitals are deploying deep learning models to support clinical work, but a model that is accurate on the day it is released does not stay accurate. Imaging devices are replaced, acquisition protocols change, and patient populations shift, so production data slowly moves away from the training data and performance degrades silently. Keeping even a small Ćeet of models healthy therefore requires a permanent team of specialists, which most hospitals do not have. Recent agentic systems, in which several language-model agents cooperate to carry out the work of a data-science team, suggest that part of this burden could be automated. This thesis examines that possibility. It Ąrst assembles the conceptual foundation the question requires, bringing together literatures that are usually treated separately: deep learning for medical imaging, the machine learning lifecycle, AutoML, MLOps, model degradation and data drift, large language models, and agentic AI. It then reviews six agentic systems published in 2025 and 2026, each analysed on the same grid of architecture, evaluation protocol, and results, and each assessed for the strength of the evidence it provides. The review yields a clear picture. These systems converge on a common shape Ů specialised agents around a language-model core, numerical optimisation delegated to deterministic tools, generated code fenced in by templates and repair loops, and memory that turns out to be load-bearing Ů and they disagree mainly on scope and on the place given to the human. Their common weakness is the same: agentic automation of model construction is advancing quickly, while the operation of deployed models is nearly untouched. Only one of the six treats drift, on synthetic data; two include no human oversight; and none ties an operational decision to a metric chosen for the clinical cost of its errors. From this gap we derive the requirements of an agentic AutoMLOps platform for hospitals: monitoring that combines immediate label-independent drift signals with delayed label-dependent performance metrics, a closed set of permitted actions, deterministic safety rules anchored on the false-negative rate, and human approval of every action that affects production Ů a design in which the language model advises and never decides. Clinova is the platform built to those requirements. It supervises a Ćeet of medicalimaging models already in service through three layers: a pipeline that prepares the data, searches the hyperparameters, and admits a new version only if its false-negative rate does not regress against the model currently in production; a controller that runs a governed observeŰthinkŰact loop over each model, detects drift on live traffic without waiting for labels, and draws every proposal from a closed set of six actions that deterministic rules conĄrm or override; and a console that gives an administrator, an engineer and a clinician three different views of the same system, and lets a clinician dispute a prediction without that dispute moving a model on its own.